application of multi-agent games to the prediction of financial time-series

Application of multi-agent games to
the prediction of financial time-series
Neil F. Johnson
a,∗
,DavidLamper
a,b
, Paul Jefferies
a
,
MichaelL.Hart
a
and Sam Howison
b
a
Physics Department, Oxford University, Oxford, OX1 3PU, U.K.
b
Oxford Centre for Industrial and Applied Mathematics, Oxford University,
Oxford, OX1 3LB, U.K.
∗
corresponding author:
Abstract
We report on a technique based on multi-agent games which has potential use in
the prediction of future movements of financial time-series. A third-party game is
trained on a black-box time-series, and is then run into the future to extract next-
step and multi-step predictions. In addition to the possibility of identifying profit
opportunities, the technique may prove useful in the development of improved risk
management strategies.
1 Introduction
Agent-based models are attracting significant attention in the study of finan-
cial markets[1]. The reasoning is that the fluctuations observed in financial
time-series should, at some level, reflect the interactions, feedback, frustration

and adaptation of the markets’ many participants (N
tot
agents).Herewere-
port on our initial results concerning the application of multi-agent games to
the prediction of future price movements[2].
Figure 1 illustrates the extent to which a multi-agent game can produce the
type of movements in price and volume which are observed in real markets.
Our game is based on the Grand Canonical Minority Game which we intro-
duced and described in earlier works[3]. Each agent holds s strategies and
only a subset N = N
0
+ N
1
of the population, who are sufficiently confident of
winning, actually play: N
0
agents choose 0 (sell) while N
1
choose 1 (buy). If
N
0
− N
1
> 0, the winning decision (outcome) is 1 (i.e. buy) and vice versa[3].
If N
0
= N
1
the tie is decided by a coin-toss. Hence N and the excess demand
Preprint submitted to Elsevier Preprint 15 May 2001

4780 4790 4800 4810
0
10
20
30
40
50
60
time
N(t)
P(t)
80
90
100
110
120
130
140
150
Fig. 1. Simulated price P (t) (solid line) and volume N(t) (bars). Here N
tot
= 101,
s =2,T = 100, r =0.53; memory m =3.
N
0−1
= N
0
− N
1
provide, to a first approximation, a ‘volume’ N(t) and ‘price-

change’ ∆P (t)attimet [3]. Here we just assume knowledge of the resulting
price-series P (t): we do not exploit any additional information contained in
N(t). Agents have a time horizon T over which virtual points are collected
and a threshold probability (‘confidence’) level r for trading. Active strate-
gies are those with a historic probability of winning ≥ r [3]. We focus on the
regime where the number of strategies in play is comparable to the total num-
ber available, and where r ∼ 0.5. In addition to producing realistic dynamical
features such as in Fig. 1, this regime yields many of the statistical ‘stylized
facts’ of real markets: fat-tailed price increments, clustered volatility and high
volume autocorrelation[3].
Exogenous events, such as external news arrival, are relatively infrequent com-
pared to the typical transaction rate in major markets - also, most news is nei-
ther uniformily ‘good’ or ‘bad’ for all agents. This suggests that the majority
of movements in high-frequency market data are self-generated, i.e. produced
by the internal activity of the market itself. The price-series P (t) can hence
be thought of as being produced by a ‘black-box’ multi-agent game whose
parameters, starting conditions (quenched disorder), and evolution are un-
known. Using ‘third-party’ games trained on historic data, we aim to generate
future probability distribution functions (pdfs) by driving these games forward
(see Fig. 2). Typically the resulting pdfs are fat-tailed and have considerable
time-dependent skewness, in contrast to standard economic models.
2
historic
price
time
probability
future price
distribution ‘corridors’
Fig. 2. Predicted distributions for future price movements.
2 Next timestep prediction

As an illustration of next timestep prediction, we examine the sign of move-
ments and hence convert ∆P(t) into a binary sequence corresponding to
up/down movements. For simplicity, we also consider a confidence threshold
level r = 0 such that all agents play all the time.
Figure 3 shows hourly Dollar $/Yen exchange-rates for 1990-9, together with
the profit attained from using the game’s predictions to trade hourly. A simple
trading strategy is employed each hour: buy Yen if the game predicts the
rate to be favourable and sell at the end of each hour, banking any profit.
This is unrealistic since transaction costs would be prohibitive, however it
demonstrates that the multi-agent game performs better than random ( ∼
54% prediction success rate). Also shown is the profit in the case when the
investment is split equally between all agents who then act independently.
Acting collectively, the N-agent population shows superior predictive power
and acts as a ‘more intelligent’ investor. As a check, Fig. 4 shows that the
game’s success returns to 50% for a random walk price-series[4].
3
70
90
110
130
150
170
1990 1992 1994 1996 1998
year
$/Yen FX
0
50
100
150
200

250
300
350
400
1990 1992 1994 1996 1998
year
profit (%)
combined population
independent agents
Fig. 3. Top: $/Yen FX-rate 1990-9. Bottom: cumulative profit for multi-agent game
(black line) and for independent agents (shaded line).
3 Corridors for future price movements
We now consider prediction over several (e.g. ten) future timesteps. As an
example, we will try to predict the large movement in Fig. 1 starting around
t = 4796. As in the case of real prices[5], it seems impossible that this drop
could have been foreseen given the prior history P (t)fort<4796. Even if
complete knowledge of the game were available, it still seems impossible that
subsequent outcomes should be predictable with significant accuracy since the
coin-toss used to resolve ties in decisions (i.e. N
0
= N
1
) and active-strategy
scores, continually injects stochasticity. We run P (t) through a trial third-
party game to generate an estimate of S
0
and S
1
at each timestep, the num-
ber of active strategies predicting a 0 or 1 respectively. Provided the black-box

game’s strategy space is reasonably well covered by the agents’ random choice
of initial strategies, any bias towards a particular outcome in the active strat-
egy set will propagate itself as a bias in the value of N
0−1
away from zero. Thus
N
0−1
should be approximately proportional to S
0
−S
1
= S
0−1
. In addition, the
number of agents taking part in the game at each timestep will be related to
the total number of active strategies S
0
+ S
1
= S
0+1
, hence the error (i.e. vari-
4
Dollar - Yen FX
45
50
55
60
012345678
time (years)

success rate %
random walk
45
50
55
60
012345678
time (years)
Dollar - Yen FX
47.9 50.3 52.6
success rate %
#players
random walk
47.5 50.1 52.6
success rate %
Fig. 4. Moving average of the multi-agent game’s success rate for the real price-series
of Fig. 3 (top left) and a random walk price-series (top right). Bottom: histogram
of individual agents’ time-averaged success rate.
ance) in the prediction of N
0−1
using S
0−1
will be approximately proportional
to S
0+1
. We have confirmed this to be true based on extensive simulations.
We then identify a third-party game that achieves the maximum correlation
between the price-change ∆P (t) and our explanatory variable S
0−1
,withthe

unexplained variance being characterized by a linear function of S
0+1
.The
predicted pdf for an arbitrary number j of timesteps into the future, is then
generated by calculating the net value of S
0−1
along all possible future routes
of the third-party game.
Figure 5 shows the ‘predicted corridors’ for P (t), generated at t = 4796 for
j = 10 timesteps into the future. Remarkably P (t) subsequently moves within
these corridors. About 50% of the large movements observed in P (t) occur in
periods with tight predictable corridors, i.e. narrow pdfs with a large mean.
Both the magnitude and sign of these extreme events are therefore predictable.
The remainder correspond to periods with very wide corridors, in which the
present method still predicts with high probability the sign of the change.
We checked that the predictions generated from the third-party game were
consistent with all such extreme changes in the actual (black-box) time series
P (t), likewise no predictions were made that were inconsistent with P (t).
5
4780 4790 4800 4810
80
90
100
110
120
130
140
150
time
P(t)

← Past Future →
P(t)
95%
75%
mean
Fig. 5. Predicted corridors for 10 future timesteps, and actual P (t)fromFig.1.The
confidence intervals and mean of the future distributions are shown.
4 Conclusion
Our initial results are encouraging. We are currently performing exhaustive
statistical studies on real financial data in order to quantify the predictive
capability of multi-agent games over different time-scales and markets.
Acknowledgements
We thank P.M. Hui, D. Challet and D. Sornette for discussions, and J. James
of Bank One for the hourly Dollar $/Yen data.
References
[1] T.LuxandM.Marchesi,Nature397, 498 (1999); M. Marsili and D. Challet,
cond-mat/0004376. See also />[2] Full details will be presented elsewhere and are the subject of a Patent
application. Supplementary material concerning the statistical tests performed
is available directly from the authors.
[3] N.F. Johnson, M. Hart, P.M. Hui and D. Zheng, Int. J. of Theor. and Appl.
Fin. 3, 443 (2000); P. Jefferies, N.F. Johnson, M. Hart and P.M. Hui, to appear
in Eur. J. Phys. B (2001); see also cond-mat/0008387.
6
[4] P. Young of Goldman Sachs has subsequently confirmed to us that these
patterns do exist in such hourly data-sets.
[5] P. Ormerod, Surprised by depression, Financial Times, February 19, 2001.
7